introduction into iup bremen’s new max-doas profile ... · introduction into iup bremen’s new...

Click here to load reader

Post on 30-Jun-2020




0 download

Embed Size (px)


  • Introduction into IUP Bremen’s new MAX-DOAS profile retrieval algorithm BOREAS Introduction into IUP Bremen’s new MAX-DOAS profile retrieval algorithm BOREAS

    Tim Bösch1, Enno Peters, Andreas Richter1, Folkard Wittrock1, Vladimir Rozanov1, and John P. Burrows1

    1 Institute of Environmental Physics (IUP), University of Bremen, Bremen, [email protected]

    1. IntroductionAlthough ground based MAX-DOAS measurements at different elevation angles have been used for several years to investigate the distribution of trace gases and aerosols, the retrieval of vertical profiles is still a difficult task and results of well-established algorithms differ strongly.Here, we introduce IUP Bremen’s new profile retrieval algorithm BOREAS (Bremen Optimal estimation REtreival for Aerosols and trace gaseS) and apply it to synthetic data computed with the radiative transfer model SCIATRAN and real data from the CINDI2 camapaign (Cabauw, 2016). The results indicate that commonly used regularization tends to give not the proper weight to the measurement which results in oscillating profiles.

    2. MAX-DOAS measurement vs. retrieval


    3. SCIATRAN [2]

    9. Acknowledgement & Selected References

    8th International DOAS Workshop 4th to 6th September 2017, Yokohama, Japan

    [1] F. Wittrock, PhD thesis, University of Bremen, May 2006[2] Rozanov, V. V., Rozanov, A. V., Kokhanovsky, A. A., and Burrows, J. P.: Radiative Transfer through Terrestrial Atmosphere and Ocean: Software Package SCIATRAN, J. Quant. Spectrosc. Ra., 133, 13–71, doi:10.1016/j.jqsrt.2013.07.004, 2014

    We thank Alexei Rozanov for his support for the IUP Bremen in-house RTM-model SCIATRAN and their ideas how to improve the approach for a further analysis. We gratefully thank European Union and ESA for their financial support for the QA4ECV and FRM4DOAS projects as well as the University of Bremen for its additional financial support.

    8. Summary Outlook

    4. Aerosol profile retrieval 5. Tikhonov regularization

    6. Trace gas retrieval

    7. CINDI2 aerosol study

    ● Azimuthal scans with different elevation angles allow retrieval of trace gases.

    ● Inverse problem is ill-imposed! ( )Adding of known information on the atmosphere (a priori information)

    ● Maximum number of degrees of freedom (DOF) is the number of altitude layers ( ). AK=I n


    Calculations within a full-spherical atmosphere including multiple scattering.

    ● Tracegas retrieval: Calculations of box air-mass factors for all geometries and altitude layers:

    ● Aerosol retrieval: Minimization of O4 optical depth of measurement and forward model by variation of the aerosol extinction profile via Tikhonov regularization.

    BAMF ij=dSCDi/VCD j



    Aer2 profile with auto scaled a-priori profileRetrieval without a-priori on a coarser vertical and spectral grid

    Why is a-priori scaling important?

    Fig. 3: Comparison of AOD (left) and bottom values (right) of retrieved profiles for different SNR and Tikhonov parameter values for aerosol scenarios without (top plots) and with prescaling of a-priori (bottom).

    Known problems● Profiles are underestimated when the

    a-priori profile is too small scaling needed● Weaker regularization leads to oscillations

    for scenarios close to the a-priori.● Chosen settings

    might be goodfor the retrieval of AOD's but bad forbottom values (or vice versa)

    50m grid, a-priori profile: exponential (1km SH, 0.18 AOD, σ: 4.0)

    Case study on the example of the elevated aerosol layer on the 12th of September 2016 in Cabauw.

    SNR: 2000, 50m grid, a-priori profile: exponential (1km SH, 0.18 AOD, σ: 4.0)12:00 13:00 14:00 15:00 16:00Standard

    Pre scaled a-priori

    ● Standard settings find a reasonable elevated layer only from 12:00 – 14:00. No convergence for 15:00 and 16:00● Pre scaled settings converge with an decreasing layer for 15:00 and 16:00h.● Bottom values and AOD are mostly more stable for Tikhonov variation with pre scaling.● RMS between measured and retrieved dSCD is smaller or similar for pre scaling compared to standard settings.

    Assumption: Variation of absorber i is equal to a scaled a-prioriThe sun-normalized intensity can be written as Taylor series:

    with retrieval parameter .

    N true , j=N apri , j⋅f j

    ln(Imeas)=ln(I apri)+ ∑absorber

    WF j⋅v j+WFaer⋅vaerf j=1+v j

    v i


    ‖ln (I apri)+WFO4,apri⋅vapri‖2→min‖ln (Imeas)+WFO 4,apri⋅vmeas‖


    Fig. 1: Representation of the aerosol weighting function, which reproduces the spectral structures of the measurement and the O4 WF well.

    This minimization problem is solved iteratively within a Thikonov regularization.x i+1=x0+(K i

    T S y−1 K i+Sr)

    −1K iT S y

    −1( y− y i+K i(x i−x0))Sr=Sa

    −1+βS tT S twith

    Step 0: DOAS FitStep 0: DOAS Fit

    DOAS fittingwith NLIN


    MAXDOAS spectraat different LOS

    Trace gas cross sections

    Step 1: aerosol retrievalStep 1: aerosol retrieval Step 2: trace gas retrievalStep 2: trace gas retrieval



    trace gas

    Cross sectionsO4

    trace gas

    BOREASoptimal estimation

    a priori information


    ● extinction coefficient profile

    ● single scattering albedo

    ● asymmetry factor● p and T profile● surface albedo

    ● concentration profile

    ● p and T profile● surface albedo

    extinction coefficient profile

    ● aerosol extinction coefficient profile● trace gas concentration profile● additional error and statistic files

    ‖WFO4 vmeas−ln(I simu)−∑WFaer vaer (zk)‖2→min

    Tikhonov regular.

    Opt. thick. of a priori aerosol

    Variation of I by a priori aerosol

    ‖y i−W i x‖2→min

    ‖τmeas−WFO4 vmeas‖2→min

    ● aerosol extinction profile● corresponding optical thickness

    ‖τmeas−τnew‖2=? min


    Other stop criteria1. dSCD RMS 2. Iterations

    BAMF calculation

    (x−xa)T Sa

    −1(x−xa)+( y−Kx)T Sϵ

    −1( y−Kx)Minimization of cost-function:

    xnew=xa+Sa KT (KS aK

    T+Sϵ)−1( y−Kxa)

    Optimal estimation:

    Forward model

    Gain matrix

    Fig. 2: Synthetic aerosol profiles and retrieved profiles for

    different parameters withand without a-priori


    β: Tikhonov parameterWF, K: weighting functionxi, yi profile and measurement vector Sa, Sy, St: Covariance/Thikonov matrix

    xnew=xapri+SaKT (KSa K

    T+g⋅Se)−1( y−Kxapri)

    A-priori pre scaling with the VCD derived from the 30° DSCD under the assumption of AMF = 1

    ● The best regularization between a-priori and measurement weighting differs strongly for individual szenarios● A-priori pre scaling improves the profiling results for days with a large variability in aerosol and trace gas

    concentrations.● Even with best settings there might be stable solutions which improve either the bottom concentration or the

    integrated concentration but not both● An automatic regularisation via derivative of rms might be a solution for optimized weighting. ● Different minima introduce a need for regularisation factor limitations by the user.

    ● A better assumption of the a-priori shape would improve the results a lot.● Improvement of the automatic RMS-based regularisation. ● Enhanced studies for "best settings" will be performed (e.g. test of settings for Gaussian profiles with

    variation of height, width and maximum value)● BOREAS will be tested for several years of measurements in Bremen to investigate the versatility of

    the algorithm for different atmospheric conditions with and without the presented pre scaling and automatization attempts.

    a-priori scaling for NO2 :

    Retrieved profiles are more stable for changes in regu-larization factors with prescaling.

    What is the perfect regu. factor?

    Fig. 4: Left: Two NO2 scenarios with different regularization factors with and without a-priori pre scaling. Bottom: VCD, bottom values and RMS between measurement and retrieved dSCD for both scenarios without a-priori pre scaling. Settings: 50m grid, a priori variance: 50%, FWHM for side diagonal elements: 0.2km

    Fig. 5: Left: DOF. Right: RMS between dSCD of measurement and retrieved profiles the derivative.

    Fig. 6: Profiles chosen from local minimum of first derivative of rms between simulation and retrieved dSCD.

    Subtraction of a polynomial

    Fig. 4

    : Ana

    lysis f


    rt for BO


    Folie 1